AM4/AM3+ Upgrade Path?...

[xSOSxHawkens]

About 3.5 years ago, I helped a buddy of mine build a new gaming system. At the time AM3+ was king of the AMD pile with the FM socket being new and dedicated to APU's.
When my friend and I discussed his core, AMD vs Intel, I explained that Intel had the raw performance lead, while AMD had better upgrade paths. When he asked what I meant, I explained that Intel generally required a new MB and core build for any new processor design, but that historically AMD had forwards and backwards compatibility, even across memory standards.

As an example, I used my (then current) system. I had an AM2 CPU (Athlon 64x2 6000+) running in forwards compatibility in an AM2+ motherboard (Foxconn A79A-S) with plans to upgrade to an AM3 CPU via backwards compatibility in the future (ended with an Athlon II x4 645). I explained that the AM3 chips were DDR3 era, but that on my AM2+ board I would be able to run even a six core Phenom with DDR2 instead. I further outlined that, since this was a multi-platform historical pattern, that he would likely be able to upgrade his CPU in the future. Historically this has been how AMD has operated for almost a decade, with compatibility between AM2/2+/3, AM2+/3/3+ and AM3/3+.

The upgrade paths are the EXACT thing that sold him on AMD, because he felt that he wanted to support the company that was supporting the consumer. He felt AMD was supporting the consumer by allowing such upgrade options vs Intel, who he saw as forcing new hardware on consumers when it was not needed.

When AMD announced the FM3 platform I was worried that they would not offer an upgrade path as they have historically done, because many outlets were pointing to FM3 as being the death punch to the AM socket series.

Now that we know via the News Announcement in this sub-reddit that there will indeed be an AM4 socket, I am left wondering, is there hope for a Zen Path on an AM3+ board (assuming motherboard vendors update the BIOS/UEFI)?

Before anyone points to DDR3 vs DDR4 as the killer point, I would counter that the AM2+/AM3 bridge saw the same change in standards from DDR2 to DDR3 and it was NOT an issue in providing backwards support.

Thoughts?

[Meurock460]

I highly doubt that AM3+ will support Zen. My guess is that the new CPUs will most likely use a different pin layout and won't physically fit in the socket. There really isn't enough info out at the moment to start talking specs on these CPUs yet though.

Honestly though, AMD is long overdue a new socket anyway. The 970 and 990 chipset and AM3+ socket has been out for nearly 5 years now. The 760g chipset found on budget AM3+ boards is even older, from the AM3 era.

Many of the motherboards lack features such as pcie 3.0 or M.2 ports, which are found on modern Intel chipsets lke Z97 or Z170.

EDIT: Even if the older socket did support Zen, wouldn't the older chipsets hold back the CPU?

[xSOSxHawkens]

I fully understand the need for a new chipset with new features, that one is a no-brainer. However, that has not hindered AMD in the past, they were able to introduce new (at the time) features for each new socket they made, while keeping pin compatibility and generational compatibility.

As for older chipsets holding cpu's back, yes it can have a minor impact, but generaly not much. My Foxconn board for example uses a 790FX chipset, the high end version of the 760g that is still in use today. The only limitaiton it has when running in forwards compatibility mode with an AM3 processor is that it uses DDR2, and rather than use the integrated memory controller on the AM3 cpu (which is DDR3) it instead uses the Chipset memory controller which is DDR2. Otherwise, there is little difference to be had.

[vadim]

xSOSxHawkens eredeti hozzászólása:

About 3.5 years ago, I helped a buddy of mine build a new gaming system. At the time AM3+ was king of the AMD pile with the FM socket being new and dedicated to APU's.
When my friend and I discussed his core, AMD vs Intel, I explained that Intel had the raw performance lead, while AMD had better upgrade paths.

OMG... Better upgrade path for the socket which was already dead that time.

xSOSxHawkens eredeti hozzászólása:

I am left wondering, is there hope for a Zen Path on an AM3+ board (assuming motherboard vendors update the BIOS/UEFI)?

I'm absolutely sure that there will not be AM3+ Zen. And, either way, why should you bother about CPU which isn't released yet? It could fail the same way as Bulldozer did.

[xSOSxHawkens]

vadim eredeti hozzászólása:

xSOSxHawkens eredeti hozzászólása:

About 3.5 years ago, I helped a buddy of mine build a new gaming system. At the time AM3+ was king of the AMD pile with the FM socket being new and dedicated to APU's.
When my friend and I discussed his core, AMD vs Intel, I explained that Intel had the raw performance lead, while AMD had better upgrade paths.

OMG... Better upgrade path for the socket which was already dead that time.

xSOSxHawkens eredeti hozzászólása:

I am left wondering, is there hope for a Zen Path on an AM3+ board (assuming motherboard vendors update the BIOS/UEFI)?

I'm absolutely sure that there will not be AM3+ Zen. And, either way, why should you bother about CPU which isn't released yet? It could fail the same way as Bulldozer did.

AM3+ was far from dead 3+ years ago. His 8350 is only about 30% slower than my 4790k clock for clock. Its more than capable of modern gaming and modern content creation. He wanted a machine for studio level audio creation with Pro-Tools, so aside from company differences, he went with more physical cores.

As for why care, becuase both he and I attempt to be forward thinking. It was part of the decision process for the platform he chose, and as such, as his friend and system builder, I am trying to keep tabs on the posibility of an upgrade when he decides he wants one.

Though, for the record, he has no interest in one for the forseeable future, as he has yet to find a single thing his computer has failed to do at a level that he accepts. From games to content creation, his machine is the fastest he has ever owned or used (aside from mine, but his GPU is twice mine, and his cpu is not that far behind).

[vadim]

xSOSxHawkens eredeti hozzászólása:

AM3+ was far from dead 3+ years ago. His 8350 is only about 30% slower than my 4790k clock for clock. Its more than capable of modern gaming and modern content creation. He wanted a machine for studio level audio creation with Pro-Tools, so aside from company differences, he went with more physical cores.

1. Name at least one AM3+ CPU released since FX-8350 (we will not count factory over/under clocked CPUs). No new CPUs for 3 years? That means socket is dead.
2. i7-4790k has about 180% FX-8350 gaming performance.
3. i7-4790k and FX-8350 have the same number of cores - 4 physical, 8 logical.

[_I_]

fx8 has 8 cores, its just that its a very poor module design, with 2 core sharing resources, making them act more like 4

zen will have its own socket, only compatable with the next gen apus
and if they keep up a design like that, its only going to gimp ther performance. again...


even if amd can magically make up for 3+ years of cpu design (and gain over 100% in core performance) they still will be behind a current skylake i5 for gaming

[vadim]

_I_ eredeti hozzászólása:

fx8 has 8 cores, its just that its a very poor module design, with 2 core sharing resources, making them act more like 4

If you wish, I can explain why FX-8/9xxx have namely 4 physical cores (they have only one branch predictor, fetch unit and decoder per module which both logical cores share the same way as Intel CPU with hyperthreading do - one cycle first thread uses shared resource, other cycle - second thread). But I'm afraid that will be offtopic in this thread.

[xSOSxHawkens]

vadim eredeti hozzászólása:

xSOSxHawkens eredeti hozzászólása:

AM3+ was far from dead 3+ years ago. His 8350 is only about 30% slower than my 4790k clock for clock. Its more than capable of modern gaming and modern content creation. He wanted a machine for studio level audio creation with Pro-Tools, so aside from company differences, he went with more physical cores.

1. Name at least one AM3+ CPU released since FX-8350 (we will not count factory over/under clocked CPUs). No new CPUs for 3 years? That means socket is dead.
2. i7-4790k has about 180% FX-8350 gaming performance.
3. i7-4790k and FX-8350 have the same number of cores - 4 physical, 8 logical.

1. I said it was far from dead 3 years ago, and it was. Three years ago it was the top socket for AMD and had new chips being released, therfore my statment stands as fact. 3 years ago it was far from dead.

2. Not even. I have a 4790k, I have tested my system litteraly side by side with his, we are best friends and I LAN at his place regualrly. When graphics are reduced to points that show CPU bottle-necking his 8350 is roughly ~30% off from my 4790k in real world performance.

I would be happy to provide numbers. For gaming we can look at GTA-V which, when set at settings to show bottle-neck on CPU result in a stock comparison (no OC) of roughly ~40fps on his rig, and 60fps on mine, a 30% difference. When I overclcok both systems (4.5ghz on his, 4.7ghz on mine) we see results of ~55fps on his and ~80fps on mine, again a roughly ~35% difference.

For content creation my system is again ~30% faster in both audio creation (Pro-Tools and audacity) as well as 3d creation (Cinnebench and POV-Ray).

For synthetic benchmarks we have again, ~30% difference in 3D-Mark (all levels) and in other synthetics like Prime-95 Benchmark and Fritz Chess Benchmark.

3. Not realy. The FX-8350 has 4 FPU cores each one tied to 2 ALU cores. This makes it an 8 core processor. It has 8 physical ALU Execution cores. Only problem is that 90% of code post-486DX FPU integration days executes on FPU's, which AMD attempts to skirt around by using FPUs as schedualers to break code out for the ALU's. Not very efficient. Does not change the fact that it still has 8 ALU cores. IN the *few* programs that execute native ALU code (mostly industrial or commercial) the FX line can still tromp an Intel (intel ALU performance is about as abysmal as AMD's FPU performance lol). Those programs are just never seen in the consumer space.

vadim eredeti hozzászólása:

_I_ eredeti hozzászólása:

fx8 has 8 cores, its just that its a very poor module design, with 2 core sharing resources, making them act more like 4

If you wish, I can explain why FX-8/9xxx have namely 4 physical cores (they have only one branch predictor, fetch unit and decoder per module which both logical cores share the same way as Intel CPU with hyperthreading do - one cycle first thread uses shared resource, other cycle - second thread). But I'm afraid that will be offtopic in this thread.

Just becuase the two ALU's share resources does not change that they are indeed two seperate cores. Saying they are not two cores becuase they chare resources is like calling a Core2Duo not dual core becuase it shares resources (to be fair, many did, I remember the whole "real dual core" thing back in the day between AMD (physically independant cores) vs Intel (seperate cores with shared cache et-all).

[vadim]

xSOSxHawkens eredeti hozzászólása:

Saying they are not two cores becuase they chare resources is like calling a Core2Duo not dual core becuase it shares resources (to be fair, many did, I remember the whole "real dual core" thing back in the day between AMD (physically independant cores) vs Intel (seperate cores with shared cache et-all).

Core 2 Duo doesn't share any resources between 2 cores, except LLC (last level cache) which isn't part of the core at all. All CPUs share LLC.
Unlike Core 2, Bulldozer shares half of pipeline stages. Namely all in-order frontend. What does this means? This means that decoder can decode up to 4 instructions and you can, conseqentally, have IPC up to 4 when core (AMD calls it "module") runs only one thread. But running second thread on the same core effectively halves IPC. No thread can execute more than 2 instructions per cycle, because it has access to decoder only every other cycle. Do you see the difference?

[vadim]

xSOSxHawkens eredeti hozzászólása:

IN the *few* programs that execute native ALU code (mostly industrial or commercial) the FX line can still tromp an Intel (intel ALU performance is about as abysmal as AMD's FPU performance lol).

FYI, even core i3 has better peak integer performance than FX-8350.

[xSOSxHawkens]

vadim eredeti hozzászólása:

xSOSxHawkens eredeti hozzászólása:

Saying they are not two cores becuase they chare resources is like calling a Core2Duo not dual core becuase it shares resources (to be fair, many did, I remember the whole "real dual core" thing back in the day between AMD (physically independant cores) vs Intel (seperate cores with shared cache et-all).

Core 2 Duo doesn't share any resources between 2 cores, except LLC (last level cache) which isn't part of the core at all. All CPUs share LLC.
Unlike Core 2, Bulldozer shares half of pipeline stages. Namely all in-order frontend. What does this means? This means that decoder can decode up to 4 instructions and you can, conseqentally, have IPC up to 4 when core (AMD calls it "module") runs only one thread. But running second thread on the same core effectively halves IPC. No thread can execute more than 2 instructions per cycle, because it has access to decoder only every other cycle. Do you see the difference?

I do, and I understand that difference (and did before you posted it). I am pretty familliar with the 2:1 ALU:FPU design present in the AMD chips, and for a long time was very excited for it. I thought that maybe just maybe we could get back to ALU code from 'le old days.

However, I still stand with the 8 core crowd. There was a time when the ALU core was the ONLY core in a cpu. Those were still considered CPU's, despite lacking an FPU core. If a 386 was a CPU with only an APU, then the 8 APU's in an FX classify it as an 8 core processor. Badly designed in a time where FPU dominates integer, but still 8 cores.

At the same time though, I get where you are comming from. For any FPU dependand code (most on the market) an FX chip is effectivly a quad-core with limited 4+ multi-thread capability.

[xSOSxHawkens]

vadim eredeti hozzászólása:

xSOSxHawkens eredeti hozzászólása:

IN the *few* programs that execute native ALU code (mostly industrial or commercial) the FX line can still tromp an Intel (intel ALU performance is about as abysmal as AMD's FPU performance lol).

FYI, even core i3 has better peak integer performance than FX-8350.

not sure where you get your numbers...

A quick google for "FX8350 Integer Performance" turns up multiple results. Here is the first...

http://www.anandtech.com/show/6396/the-vishera-review-amd-fx8350-fx8320-fx6300-and-fx4300-tested/2

As you can see, in the Integer dependant 7-Zip compression test, the 8350 beats out the i7-3770k, and in fact ALL bulldozer parts beat out their equivelent Ivy Bride counterparts in the ineger department...

Again, in this review they use the DIEP Chess Benchmark, a program which uses NO FPU code at all, and is entirely dependant on Integer performance, and again we see the 8350 beat out an i7...

http://www.extremetech.com/computing/138394-amds-fx-8350-analyzed-does-piledriver-deliver-where-bulldozer-fell-short/2

And finaly, yet again, in productivity apps which use Integer Performance as a main under-pin we can see the 8350 killing the i5 sectors and going toe to toe with i7's and beating multiple models of them...

http://techreport.com/review/23750/amd-fx-8350-processor-reviewed/10

Mind providing a source showing an i3 beating an 8350 in integer, because I cant find one (but would be happy to read it).

[vadim]

xSOSxHawkens eredeti hozzászólása:

However, I still stand with the 8 core crowd. There was a time when the ALU core was the ONLY core in a cpu. Those were still considered CPU's, despite lacking an FPU core. If a 386 was a CPU with only an APU, then the 8 APU's in an FX classify it as an 8 core processor. Badly designed in a time where FPU dominates integer, but still 8 cores.

At the same time though, I get where you are comming from. For any FPU dependand code (most on the market) an FX chip is effectivly a quad-core with limited 4+ multi-thread capability.

No. I didn't say anything about floating-point ops at all. I said that threads shares part of pipeline and running second thread affect first thread performance. Very similar to Intel and future Zen SMT. Only difference is SMT logical cores share all pipeline stages while CMT logical cores share only first half of pipeline.

[xSOSxHawkens]

vadim eredeti hozzászólása:

xSOSxHawkens eredeti hozzászólása:

However, I still stand with the 8 core crowd. There was a time when the ALU core was the ONLY core in a cpu. Those were still considered CPU's, despite lacking an FPU core. If a 386 was a CPU with only an APU, then the 8 APU's in an FX classify it as an 8 core processor. Badly designed in a time where FPU dominates integer, but still 8 cores.

At the same time though, I get where you are comming from. For any FPU dependand code (most on the market) an FX chip is effectivly a quad-core with limited 4+ multi-thread capability.

No. I didn't say anything about floating-point ops at all. I said that threads shares part of pipeline and running second thread affect first thread performance. Very similar to Intel and future Zen SMT. Only difference is SMT logical cores share all pipeline stages while CMT logical cores share only first half of pipeline.

I got that part, that is effectivly what causes the FPU bottle-neck though. :)

As shown above, in ALU dependant code, the 8350 can still rock it hard. It just has to have native Integer code.

infact, using Anandtech as a refferance, since they use the same 7-Zip Integer benchmarks, we can see how the new Skylake 6700k compares in iteger performance vs an 8350.

In the 7-Zip bench, with 32mb Dictionary size the total MIPS are:

FX-8350 : 23,407
i7-6700k : 26,370

Thats only an 11% improvment for a top of the line brand new i7 vs a multi-year old low clocked bulldozer core. As I said, Intel Integer performance is verry poor compared to their FPU. Basically the inverse of AMD lol...


Edit for source,

FX article is first one listed above, i7 is here
http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/11

[vadim]

xSOSxHawkens eredeti hozzászólása:

As you can see, in the Integer dependant 7-Zip compression test, the 8350 beats out the i7-3770k, and in fact ALL bulldozer parts beat out their equivelent Ivy Bride counterparts in the ineger department...

Well, we say about completely different things. I told about peak performance. It very easy to calculate.
AMD core (module) has 4 64-bit ALUs. Each of them capable to execute 1 integer op per cycle.
I.e. each core can run up to 4 ops per cycle. FX-8350 has 4 modules and can run up to 16 ops per cycle.
Intel core also has 4 ALUs. But 3 of them are 256-bit and can run 4 integer ops per cycle when running AVX 2 code. 4th ALU is 64-bit and cannot run vector instruction. So, each core can run 3*4+1=13 integer instructions per cycle. i3 has 2 cores which give us 26 integer ops per cycle total.
As you know, FX has no hardware AVX support (its FPU registers are only 128-bit) and no AVX 2.0 support at all.
But 7zip doesn't use AVX2.0 and core i3 couldn't get advantage from it. Because of that real performance is much less than theoretical peak performance.